基于逐层扩张卷积的3D牙颌实例分割  

3D dental instance segmentation with graph layer-wise dilated convolutions

作　　者：唐瑞成 成苗 何莲 石向文 郭静 赵铱民 王胜朝 余快 TANG Ruicheng;CHENG Miao;HE Lian;SHI Xiangwen;GUO Jing;ZHAO Yimin;WANG Shengchao;YU Kuai(Chengdu Institute of Computer Application,Chinese Academy of Sciences,Chengdu Sichuan 610041,China;School of Computer Science and Technology,University of Chinese Academy of Sciences,Beijing 100049,China;Shenzhen CBPM-KEXIN Banking Technology Company Limited,Shenzhen Guangdong 518206,China;State Key Laboratory of Oral&Maxillofacial Reconstruction and Regeneration,Xi’an Shaanxi 710032,China;National Clinical Research Center for Oral Diseases,Xi’an Shaanxi 710032,China;Shaanxi Key Laboratory of Stomatology,Xi’an Shaanxi 710032,China;School of Stomatology,The Fourth Military Medical University,Xi’an Shaanxi 710032,China;School of Basic Medical Science,The Fourth Military Medical University,Xi’an Shaanxi 710032,China)

机构地区：[1]中国科学院成都计算机应用研究所,成都610041 [2]中国科学院大学计算机科学与技术学院,北京100049 [3]深圳市中钞科信金融科技有限公司,广东深圳518206 [4]口颌系统重建与再生全国重点实验室,西安710032 [5]国家口腔疾病临床医学研究中心,西安710032 [6]陕西省口腔医学重点实验室,西安710032 [7]第四军医大学口腔医院,西安710032 [8]第四军医大学基础医学院,西安710032

出　　处：《计算机应用》2024年第S01期235-241,共7页journal of Computer Applications

摘　　要：在计算机辅助的正畸手术规划和牙齿识别中,准确地从3D牙颌模型中分割出每颗牙齿至关重要。无需候选区域的分割方法虽然高效且易用,但存在诸多限制:在局部特征聚合时,现有方法依赖欧氏距离寻找特征空间内最近的多个网格单元,只能实现较小的感受野,这可能导致对牙齿网格的错误预测;此外,现有方法未充分考虑3D牙颌模型的方向性和同一牙齿实例的网格单元间的潜在信息。针对上述问题,提出一种名为“逐层扩张卷积”的卷积方法,它能在有效提取特征的同时,不增加任何计算成本;同时,提出方向特征对齐模块和潜在信息挖掘模块,以进一步优化牙颌分割结果。在真实采集的3D牙颌内口扫描数据集上的实验结果表明,所提方法在总体准确率(OA)和平均交并比(mIoU)方面分别达到了98.46%和94.47%,相较于TSGCNet(Two-Stream Graph Convolutional Network),分别提升了3.27个百分点和10.45个百分点。Accurate segmentation of each tooth from a 3D dental model is essential in computer-aided planning of orthodontic surgery and identity recognition with teeth.Proposal-free methods are efficient and convenient,but there are many limitations.Existing methods directly find the multiple nearest cells with the smallest Euclidean distance in feature space for local feature aggregation with a small receptive field,which can lead to some false predictions.In addition,the existing methods ignore the 3D dental model’s directionality and potential information among cells from the same instance.To address these issues,an innovative method called graph layer-wise dilated convolution was proposed to extract local features with a larger receptive field without increasing the computational cost.Furthermore,a directional features alignment module and potential information mining module were proposed to improve the instance segmentation results.Experimental results on a real 3D dental scan dataset show that compared to TSGCNet(Two-Stream Graph Convolutional Network),the proposed method achieves 98.46%and 94.47%in terms of Overall Accuracy(OA)and mean Intersection over Union(mIoU)with improvements of 3.27 percentage points and 10.45 percentage points,respectively.

关 键 词：3D牙颌实例分割 计算机辅助规划 感受野 局部特征聚合 牙颌方向性 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]